The Designware FIFO is filled in units of 32 bit words, but the buffer
we are passed is not guaranteed to be a multiple of 4 bytes long, nor
aligned on a 4-byte boundary. So in theory we are reading 0-3 bytes
of unused garbage from the end of the array.
That's currently benign on supported platforms with this hardware,
which all support misaligned reads. But not all do. And the incoming
arrival of memory protection opens the possibility that those extra
bytes would cross a protection boundary and cause a crash or security
bug.
Do this right.
(Note that this is fixed to little endian byte order. The Designware
databook is frustratingly silent on the endianness it expects, but
existing hardware I can see is definitely LE and I see a few spots in
the Linux dwc2 driver that likewise assume LE).
Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
eebb17ca52